Graphics chips could draw but not compute, until the Xbox arrived.
Console graphics came from bespoke, fixed-function chips for more than two decades, until the original Xbox burst onto the scene in 2001 with a PC GPU inside. Those of us who played on early consoles like the NES will have come across some familiar quirks with those fixed-function chips: When too many Goombas lined up on the same row in Super Mario Bros., for example, some of them would start to flicker. That’s the NES’s graphics chip hitting its hard limit of eight sprites per scanline; the chip drops anything past eight, and the game alternates which ones.
In the NES, it was the Ricoh 2C02 Picture Processing Unit that rendered game graphics. A pretty rudimentary chip, it couldn’t run any code. Instead, it took tile artwork from the cartridge and a couple of lookup tables from memory and turned them into a TV signal, one scanline at a time. That arrangement — a dedicated picture generator fed by the CPU — is how most consoles rendered their graphics from the late 1970s into the 2000s.
When Microsoft’s first Xbox came out in November 2001, it didn’t look much like any console before it. Instead, it was built around an Intel Pentium III and a graphics chip NVIDIA made specially for it, based on the GeForce 3. The GeForce 3 had only been announced that February, and it was the first GPU that let developers program their own pixel and vertex shaders.
Sprites and scanlines
The NES’ Ricoh 2C02 worked from about 10kB of memory. That’s broken down into roughly 8kB of tile shapes stored on the cartridge and 2kB of RAM in the console holding a background map or two. Separate small memories inside the chip tracked things like the position, shape and orientation of up to 64 sprites. Its output was 256×240 pixels, built from 8×8 tiles.
As the CRT TV’s electron beam swept across each line of the screen, the 2C02 worked out which 8×8 tile belonged at each spot. It then cross-referenced this with which sprites were sat on that line and decided which pixel to show. This took place in step with the beam, so the CPU couldn’t safely rewrite the picture while it was being drawn.
Unlike the NES, the earlier Atari 2600 made the CPU also handle graphics. Because its Television Interface Adaptor had no frame buffer, the processor had to feed new values to the chip for every scanline drawn by the beam. Just five years earlier in 1972, the Magnavox Odyssey had no processor or memory, period, drawing its dots and lines with discrete diode-transistor logic. All of these machines were designed around a CRT, which is partly why retro games look better on old CRT TVs.
The PlayStation 2’s custom silicon shared the load
Sony showed off the PS2’s chips in March 1999. The CPU was the Emotion Engine, a 128-bit design Sony built with Toshiba, and graphics went to a separate chip called the Graphics Synthesizer. In the final consoles, the two ran at about 295 MHz and 147 MHz.
The Graphics Synthesizer didn’t do any geometry work, though. By the time a triangle reached it, it had already been placed within the frame, and all the chip had to do was fill it in with color and texture. The geometry work happened back on the Emotion Engine, in two vector units. The CPU could give VU0 odd jobs, while VU1 mostly ran the 3D transforms that turned a game’s models into on-screen triangles. Sony and Toshiba’s engineers said a single vector unit could manage 85 million of those transforms a second. While no two consoles were the same, many early 3D consoles similarly shared the graphical load across multiple chips.
Interestingly, Sony had used the term GPU for the 1994 PlayStation’s Toshiba-designed graphics chip. That’s five years before NVIDIA marketed the GeForce 256 as the world’s first GPU. But the working definition of a modern GPU, a chip that runs small programs written by the game, only came with GeForce 3 in early 2001.
The original Xbox and everything after
NVIDIA and Microsoft signed their Xbox deal in March 2000, and five days later, Bill Gates took to the stage and announced the console. It would use an Intel Pentium III-class processor, a custom NVIDIA graphics processor, 64MB of memory shared by both and an 8GB hard drive. The finished box was unveiled at CES 2001 and then went on sale that November.
But consoles didn’t make the switch to PC-style hardware all at once. The Nintendo GameCube launched the same month with a PowerPC processor from IBM, and every major console used one for the next decade: the Xbox 360 dropped Intel for three PowerPC cores from IBM in 2005, the PS3’s Cell chip was built by Sony, Toshiba and IBM around a PowerPC core and the Wii and Wii U both ran updated versions of the GameCube’s IBM chip.
The graphics side had already gone PC-style, with the 360’s ATI chip and the PS3’s NVIDIA one, but the processors hadn’t. That changed in 2013, when the PS4 moved to eight AMD processor cores and a graphics design taken from AMD’s Radeon PC cards. Today’s PS5 and Xbox Series X each run eight AMD Zen 2 cores alongside AMD RDNA 2 graphics on the same chip as the CPU. Ultimately, designing graphics chips from scratch for consoles stopped making sense once off-the-shelf GPU architectures fit a console’s budget.
